Principles
Five rules, each one a behavior you can observe.
Brisora is built to fail closed. When it cannot establish something, it says so and stops, rather than proceeding on a best guess. Five rules follow from that, and none of them makes the engine more permissive.
Deterministic evaluation
The same evidence, policy and evaluation time always produce the same outcome.
- No language model runs anywhere in the decision path
- The evaluation time is supplied explicitly; rules do not read the clock
- Monetary arithmetic uses whole minor units, so totals do not drift
Evidence before approval
A fact counts only when it resolves to an exact passage of a retained source.
- An evidence reference that does not match the retained content exactly invalidates the reading
- A mandate can require specific evidence for an action; if it is missing, the outcome is ESCALATE, never ALLOW
- A tender notice is qualified only on what the notice itself evidences
Unknown is not an assumption
What cannot be established stays explicit and blocks the result that depends on it.
- An unknown cost component is never counted as zero; a required one makes the landed cost UNKNOWN with no total
- A missing exchange rate is never replaced by a default
- An action class the engine does not recognize is denied
- Conflicting values in a notice are both kept; neither is chosen
Authority cannot expand itself
A commercial mandate is issued by people and is read-only to the engine.
- Authority is an allow-list: what a mandate does not name is denied
- The engine has no function that creates, edits or activates a mandate
- A mandate seen in one request confers nothing on the next
Human review cannot self-approve
ESCALATE names what a person must decide. The engine cannot supply that decision.
- An escalated evaluation stays escalated however often it is asked again
- No input field can carry an approval into an evaluation
- Brisora MCP has no tool that approves anything
Where these show up
Enforced in the engine, visible in every result.
The first and third rules shape landed cost and commercial-policy evaluation. The second shapes tender qualification and its evidence. All five are why a result carries its own decision provenance, and the system architecture is what keeps them out of reach of the AI client.